Stroke is a leading cause of mortality and long term disability, and early prediction using machine learning models can support clinical decision making. In this study, a hybrid, stable, and interpretable feature selection framework is employed together with the LightGBM algorithm for stroke prediction. First, SHAP values derived from an XGBoost model are used to quantify the importance of each feature. Then, the absolute correlation of each feature with the target variable stroke and its stability across cross validation folds (i.e., the number of times it appears among the top 10 features) are considered as complementary criteria. These three measures are combined into a weighted composite score, and the top 10 features are selected for model training. The results show that, although some previous studies report higher overall accuracy, the proposed model achieves a Recall of 100% in identifying stroke patients, along with a Precision of 97.12% and an appropriate F1 score, providing a favorable balance between reducing missed cases and limiting false alarms.